Collecting Food and Drink Intake Data With Voice Input: Development, Usability, and Acceptability Study

**Background:** Epidemiological cohorts typically collect self-reported data at widely spaced intervals (every 1-5 years), which is inadequate for traits that vary acutely (e.g., dietary intake, mental well-being). Voice-based systems like Amazon Alexa could enable real-time, continuous self-report data collection with lower burden than traditional diaries or web-based recall tools. This pilot study aimed to demonstrate technical feasibility, assess participant acceptability, and provide an initial evaluation of data validity using food and drink diaries as an exemplar.

**Methods:** 45 volunteers were recruited from University of Bristol staff and student email lists. Participants were asked to use an Amazon Alexa custom skill to report what they ate or drank for 7 days, and to also submit the same information via a web-based form. The Alexa skill allowed participants to add date/time of intake, add items, cancel, and submit events. Equipment was delivered to participants' homes due to the COVID-19 pandemic. After the study period, participants completed a postparticipation questionnaire on usability and acceptability. A separate nonparticipation questionnaire was sent to those who did not take part. Counterpart entries (Alexa and web entries corresponding to the same intake event) were identified using intake and submission timestamps. Food and drink descriptions were compared using both an automated word-set comparison and a systematic manual review by the principal investigator (LACM) with inter-rater reliability checks by 5 independent researchers.

**Key Results:** Of 45 registered participants, 37 (82%) comprised the analytical sample; 29 participants had both Alexa and web entries that could be matched (comparison sample). Most participants were aged 20-39 years (30/37, 81%) and female (23/37, 62%). Participants completed more web diary entries than Alexa entries (median 17, IQR 13-27 vs median 11, IQR 7-21; paired t-test P<.001). The median number of partial (unsuccessful) Alexa attempts was 6 (IQR 1-9). Across 310 counterpart entries, 71.6% (222/310) had matching intake timestamps. Of 588 Alexa items reviewed manually, 357 (60.7%) contained the same food/drink information as the corresponding web entry. Among the 194 items with differences: 64 (33%) had less detail from Alexa, 12 (6.2%) had more detail, 15 (7.7%) had different detail, 36 (18.6%) had an Alexa entry issue, and 59 (30.4%) had a misspelling in Alexa (40/59, 68% due to Alexa recording "to" instead of "two"). 28 Alexa items (5%) had a major entry issue. Alexa interjected often or always when participants were telling her what they ate/drank (18/35, 51%) and less often for date/time (7/35, 20%). Despite this, 26/35 (74%) said they would be happy to use a voice-controlled system at home for future research, and 28/35 (80%) would use one on a wearable device. Among 69 nonparticipants, 11 (16%) cited privacy concerns; 42 (61%) said they would be happy to use Alexa at home for future research.

**Clinical Implications:** Voice-based data collection is technically feasible for epidemiological research but requires significant improvements in interface design. The conversational interface was a major source of frustration, causing interruptions and incomplete entries. A less conversational design (e.g., allowing participants to state all information without separate prompts) and integration with a phone app for validation could improve usability and data quality. Voice-based approaches may be particularly valuable for populations with difficulty writing (e.g., dyslexia, motor neuron disease). Further studies are needed to compare different voice systems (e.g., Google Assistant), device types (wearables vs home devices), and to evaluate biases in collected data before this approach can be deployed in large-scale epidemiological studies.